Working w/Claude Code

AI orchestration. Instead of writing a script (or a class) to do something, think of all your architecture as either 1) an AI agent or 2) a tool that feeds an agent. Context passed along across graph nodes. LLM reasoning within each step “do I have everything I need? Or should I use one of my tools?” And a supervisor agent runs above everything. Evaluates each agent’s task and decides whether it needs to redo its work (with a new prompt or tooling) or pass it along to the next stage. Constantly QA’ing the agentic workflow and A/B testing different prompts to see what improves things. Solving your problem as if it was designing a neural network (without the complex math) — derivatives and linear algebra. Bayesian statistics.

Drastic difference working on Windows versus MacOS. Claude so much slower on Windows since it has to interface with the operating system through PowerShell. Too many quirks in that encapsulation. Prefers the Unix based OS running on Mac. Writes code faster.

AI can logic and reason better than most humans nowadays. And thereby replaces those fields at scale. Why academics have been arguing that Classics and Neuroscience will be those best positioned in the AI age. Human-to-AI differentiation is nuanced and difficult to formalize.